Voting Data-Driven Regression Learning for Accelerating Discovery of Advanced Functional Materials and Applications...

Area of Science:

• Materials Science
• Machine Learning
• Condensed Matter Physics

Background:

• Regression machine learning models are crucial for predicting material properties.
• Performance is often limited by insufficient available materials data.
• Accurate prediction of ferroelectric properties is vital for advanced electronic devices.

Purpose of the Study:

• To develop a novel voting data-driven method to enhance regression learning model performance for materials property prediction.
• To apply this method to identify novel two-dimensional (2D) hexagonal binary ferroelectric compounds.
• To extract actionable insights into factors influencing ferroelectric polarization.

Main Methods:

• Development of a voting data-driven ensemble method for regression.
• Application to a dataset of 2135 two-dimensional hexagonal binary compounds.
• Utilizing unsupervised learning to analyze atomic and electronic descriptors influencing polarization.

Main Results:

• Significant improvement in the performance of the regression model for predicting electric polarization.
• Screening of 38 stable ferroelectrics with out-of-plane polarization (31 metals, 7 semiconductors).
• Identification of key atomic factors (valence electrons, ionic polarizability, electronegativity) affecting polarization.

Conclusions:

• The voting data-driven method effectively enhances materials property prediction accuracy, even with limited data.
• This approach facilitates the discovery of new functional materials, specifically ferroelectrics.
• Unsupervised learning provides valuable insights into structure-property relationships for targeted material design.